Built an automated resource management system using Kubernetes, gCloud, Python, Docker, and Bash that dynamically schedules and allocates resources between latency-sensitive and batch workloads, ensuring SLO compliance while maximizing throughput and system stability.
Built and evaluated deep learning models (LSTM, BiLSTM, Transformer) on pre-processed ICU time-series data from PhysioNet for mortality prediction, using representation learning and visualization techniques (t-SNE, UMAP) for clinical interpretability.
Built and compared machine learning models (Logistic Regression, Random Forest, XGBoost, Neural Networks) on processed clinical data to predict heart disease risk, using SHAP values and evaluation metrics to improve accuracy and interpretability.
Fine-tuned transformer-based LLMs (BERT, RoBERTa, DeBERTa v3, XLM-R) on a balanced 100K+ sentence dataset for ternary sentiment classification, using ensemble methods like softmax averaging and majority voting to boost robustness and performance.